Add support for transformers v5.5

[echarlaix]

Adds compatibility with transformers v5.5 and bumps the supported range to >=4.57,<5.6.

• attention_mask is now created explicitly when needed, since generate() no longer creates it since v5.2
• mm_token_type_ids is now computed and passed to get_rope_index as required by transformers v5.3 for Qwen2VL, Qwen2.5VL, and Qwen3VL models
• position_ids generation fixed for v5.3 decode step where input_ids is already sliced to 1 token
  • cache_position handling for SSM models (>=v4.54): SSM models no longer receive cache_position from generate(). It is now inferred from cache_params when absent
• AutoModelForVision2Seq → AutoModelForImageTextToText (>=v4.54)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[echarlaix]

failing in v5.2 https://github.com/huggingface/transformers/blob/v5.2.0/src/transformers/modeling_utils.py#L2315 fixed in v5.3 since huggingface/transformers@a64996e

[regisss]

LGTM

[rkazants]

please clean-up preview models (gemma 4, qwen3.5) job because they should come to regular jobs

[IlyasMoutawwakil]

LGTM !

[rkazants]

We loose support for critical models:
minicpm | MiniCPMOpenVINOConfig | 4.53.3 | 469
minicpm3 | MiniCPM3OpenVINOConfig | 4.53.3 | 478
chatglm | ChatGLM2OpenVINOConfig | 4.55.4 | 518
qwen | QwenOpenVINOConfig | 4.55.4 | 822
arctic | ArcticOpenVINOConfig | 4.53.3 | 1219
llava-qwen2 | LlavaQwen2OpenVINOConfig | 4.53.3 | 2133
minicpmo | MiniCPMOOpenVINOConfig | 4.51.3 | 2752
phi3_v | Phi3VisionOpenVINOConfig | 4.53.3 | 2769
phi4_multimodal | Phi4MMOpenVINOConfig | 4.53.3 | 2904
deepseek / deepseek_v2 | DeepseekOpenVINOConfig | 4.53.3 | 391

[rkazants]

may be let us leave ["4.45.0", "4.57.6", "latest"]?

[echarlaix]

@rkazants we only test min/max supported transformers version (4.57.6 and 5.5)
https://github.com/huggingface/optimum-intel/blob/transformers-v5.5/.github/workflows/test_openvino.yml#L42 having more tests can result in rate limit errors as this launches x8 tests each time (can be an issue especially for concurrent PRs), would make more sense to have nightly tests for missing transformers release